Applied LLM Engineering with TypeScript

A comprehensive, project-based course teaching applied LLM engineering from first API calls through production deployment. Built in TypeScript with the Vercel AI SDK.

How It Works

• 27 modules across 6 parts, ~198-251 hours
• Multi-provider — learn with Mistral (default, free tier), Groq, Anthropic, OpenAI, or Ollama (local)
• Vercel AI SDK — provider-agnostic patterns that work everywhere
• Gamification — earn XP, unlock badges, climb ranks from Token to LLM Architect
• Claude Code guided — each module taught section-by-section with interactive exercises

Getting Started

bun install
cp course/preferences.example.toml course/preferences.toml
# Edit preferences.toml to match your background and provider

To start a module, run the corresponding Claude Code command: /module-1

To check progress: bun run progress

Curriculum

Part I: First Contact

#	Module	Topics	Hours
1	Setup & First LLM Calls	Bun, AI SDK, providers, generateText, streamText, roles, temperature	5-7
2	Prompt Engineering	System prompts, few-shot, chain-of-thought, templates, versioning	5-7
3	Structured Output	Output.object, Zod schemas, type-safe responses, validation	5-6

Part II: Core Patterns

#	Module	Topics	Hours
4	Conversations & Memory	Multi-turn, context windows, sliding window, summarization	6-8
5	Long Context & Caching	Prompt caching, KV cache, context compression, chunked prefill	6-8
6	Streaming & Real-time	streamText, Output.object, SSE, backpressure, UI patterns	6-8
7	Tool Use	Zod tool definitions, execution, multi-step loops, stopWhen	7-9
8	Embeddings & Similarity	Embedding models, cosine similarity, vector stores, semantic search	7-9
9	RAG Fundamentals	Chunking, retrieval, context injection, citation, pipeline	8-10

Part III: Advanced Retrieval

#	Module	Topics	Hours
10	Advanced RAG	Hybrid search, reranking, HyDE, self-RAG, evaluation	8-10
11	Document Processing	PDF/HTML/markdown extraction, recursive chunking, metadata	8-10
12	Knowledge Graphs	Entity extraction, relationship mapping, graph RAG	8-10
13	Multi-modal	Vision models, image understanding, audio, multi-modal RAG	8-10

Part IV: Agents & Orchestration

#	Module	Topics	Hours
14	Workflows & Chains	Sequential/parallel pipelines, branching, composable chains	7-9
15	Durable Workflows	Workflow SDK: use step/use workflow, sleep, hooks, retries, WorkflowAgent	8-10
16	Agent Fundamentals	ReAct pattern, planning loops, tool selection, observation cycles	8-10
17	Multi-Agent Systems	Orchestrator-worker, delegation, shared state, communication	7-9
18	Eve Fundamentals	Filesystem-first agents: tools, skills, approval, model config, eve eval	8-10
19	Eve in Production	Channels, schedules, subagents, connections, durability, deploy	8-10
20	Code Generation	LLM-generated code, sandboxed execution, iterative refinement	8-10
21	Human-in-the-Loop	Approval flows, feedback integration, active learning	7-9

Modules 15, 18, 19 teach the Vercel Workflow SDK and Eve frameworks in isolated apps/ mini-projects (Node ≥24 for Eve; @workflow/vitest / eve eval harnesses), not the src/ + bun test flow.

Part V: Quality & Safety

#	Module	Topics	Hours
22	Evals & Testing	LLM-as-judge, benchmarks, regression suites, prompt CI	8-10
23	Fine-tuning	When to fine-tune, dataset prep, training pipelines, evaluation	8-10
24	Safety & Guardrails	Input validation, output filtering, jailbreak prevention	8-10
25	Cost Optimization	Semantic caching, model routing, token budgets, fallback chains	8-10

Part VI: Production

#	Module	Topics	Hours
26	Observability	Logging, tracing, token tracking, latency metrics, debugging	8-10
27	Deployment	Hono server, auth, rate limiting, scaling, provider failover	10-12

Estimated Time

Part	Modules	Hours
I. First Contact	1-3	15-20
II. Core Patterns	4-9	40-52
III. Advanced Retrieval	10-13	32-40
IV. Agents & Orchestration	14-21	61-77
V. Quality & Safety	22-25	32-40
VI. Production	26-27	18-22
Total	1-27	~198-251

Prerequisites

• TypeScript (comfortable with types, generics, async/await)
• Terminal comfort (running commands, environment variables)
• API key for at least one provider (Anthropic, OpenAI, Mistral, Groq) OR Ollama installed locally
• No AI/ML background required — we start from first principles

Module Dependencies

• Parts I-III are sequential (each builds on the previous)
• Parts IV and V can be done in either order after Part III
• Part VI requires all prior parts